Supernova Classification using the Recurrent Neural Network in the CSST Ultra-Deep Field Survey

TL;DR

This work tackles the challenge of extracting robust cosmological constraints from photometric SN surveys by combining a recurrent neural network classifier (SuperNNova) with SALT3-based SN Ia modeling and JLA-like cuts in the CSST-UDF context. Mock light curves for SNe Ia and CCSNe are generated with SNCosmo, enabling a two-stage classification and selection workflow that yields a high-purity SN Ia sample (about 2{,}197 Ia with only 4 CCSN contaminants after cuts). The cosmology is then inferred using BEAMS with Bias Corrections (BBC) in a two-step binning framework, with an MCMC fit to flat $w$CDM parameters, delivering constraints of $\Omega_{\rm M}=0.304^{+0.030}_{-0.052}$ and $w=-1.017^{+0.177}_{-0.189}$ (1$\sigma$), demonstrating that photometric data alone can achieve competitive precision. Overall, the study shows that the CSST-UDF, when paired with sophisticated RNN-based classification and BBC-based bias corrections, can reliably probe the expansion history of the Universe, offering a scalable path for future Stage IV surveys.

Abstract

We study supernova (SN) classification using the machine learning method of the Recurrent Neural Network (RNN) in the Chinese Space Station Survey Telescope Ultra-Deep Field (CSST-UDF) photometric survey, and explore the improvement of the cosmological constraint. We generate the mock light curve data of Type Ia supernova (SN Ia) and core collapse supernova (CCSN) using SNCosmo with SALT3 SN Ia model and CCSN templates, and apply the SuperNNova (SNN) program for classifying SNe. Our study indicates that the SNN combined with the Joint Light-curve Analysis like (JLA-like) cuts can enhance the purity of the CSST-UDF SN Ia sample up to over 99.5% with 2,193 SNe Ia and 4 CCSNe, which can significantly increase the reliability of the cosmological constraint results. The method based on the Bayesian Estimation Applied to Multiple Species (BEAMS) with Bias Corrections (BBC) framework is used to correct the SN Ia magnitude bias caused by the selection effect and CCSN contamination, and the Markov Chain Monte Carlo (MCMC) method is employed for cosmological constraints. We find that the accuracy of the constraints on the matter density $Ω_{\rm M}$ and the equation of state of dark energy $w$ can achieve 14% and 18%, respectively, assuming the flat $w$CDM model. This result is comparable to that from the current surveys that relied on spectroscopic confirmation. It indicates that our data analysis method is effective, and the CSST-UDF SN photometric survey is powerful in exploring the expansion history of the Universe.

Figures (6)

• Figure 1: The mock light curve examples for SNe Ia in different CSST-UDF photometric bands at redshifts between $z = 0.28$ and 1.3. The solid lines correspond to the theoretical expectations derived from the fiducial model.
• Figure 2: The mock light curve examples at $z \simeq 0.5$ for the six types of CCSNe considered in this study. The solid lines denote the theoretical light curves derived from the fiducial model.
• Figure 3: A simplified architecture of an LSTM network. The cell state ($C_t$) carries information through time. The input layer receives data at each time step, while hidden layers compute and update the hidden state ($H_t$). The output layer generates predictions. The diagram illustrates the flow of information across time steps $t \in [1, 4]$.
• Figure 4: An LSTM cell within the hidden layer, comprising the cell state ($C_t$), hidden state ($H_t$), and input ($X_t$). It includes key components, i.e. Input Gate, Forget Gate, and Output Gate, which enable the model to effectively capture long-term dependencies.
• Figure 5: The Hubble diagram as a function of input redshifts for the 2197 SNe which are classified by the SNN. The blue and orange data points denote SNe Ia and CCSNe, respectively. In the lower panel, we show the residuals of the distance modulus and errors relative to the fiducial cosmology for the 24 redshift bins.